US10496730B2 - Factor analysis device, factor analysis method, and factor analysis program - Google Patents

Factor analysis device, factor analysis method, and factor analysis program Download PDF

Info

Publication number
US10496730B2
US10496730B2 US15/125,302 US201415125302A US10496730B2 US 10496730 B2 US10496730 B2 US 10496730B2 US 201415125302 A US201415125302 A US 201415125302A US 10496730 B2 US10496730 B2 US 10496730B2
Authority
US
United States
Prior art keywords
time series
feature
explanatory
series
time
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Active, expires
Application number
US15/125,302
Other languages
English (en)
Other versions
US20170109324A1 (en
Inventor
Takehiko Mizoguchi
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
NEC Corp
Original Assignee
NEC Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by NEC Corp filed Critical NEC Corp
Assigned to NEC CORPORATION reassignment NEC CORPORATION ASSIGNMENT OF ASSIGNORS INTEREST (SEE DOCUMENT FOR DETAILS). Assignors: MIZOGUCHI, TAKEHIKO
Publication of US20170109324A1 publication Critical patent/US20170109324A1/en
Application granted granted Critical
Publication of US10496730B2 publication Critical patent/US10496730B2/en
Active legal-status Critical Current
Adjusted expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
    • G06F17/10—Complex mathematical operations
    • G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
    • G06K9/00496—
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • G06N5/045—Explanation of inference; Explainable artificial intelligence [XAI]; Interpretable artificial intelligence
    • G—PHYSICS
    • G07—CHECKING-DEVICES
    • G07C—TIME OR ATTENDANCE REGISTERS; REGISTERING OR INDICATING THE WORKING OF MACHINES; GENERATING RANDOM NUMBERS; VOTING OR LOTTERY APPARATUS; ARRANGEMENTS, SYSTEMS OR APPARATUS FOR CHECKING NOT PROVIDED FOR ELSEWHERE
    • G07C3/00—Registering or indicating the condition or the working of machines or other apparatus, other than vehicles
    • G07C3/14—Quality control systems
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F2218/00—Aspects of pattern recognition specially adapted for signal processing
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/02—Total factory control, e.g. smart factories, flexible manufacturing systems [FMS] or integrated manufacturing systems [IMS]
    • Y02P90/22—

Definitions

  • the present invention relates to a factor analysis device, a factor analysis method, and a factor analysis program, and particularly, relates to a factor analysis device, a factor analysis method, and a factor analysis program that identify an explanatory time series that has an influence on a change in value of a response time series.
  • a factor identification method that identifies an explanatory time series that influences a change in value of a response time series is used in a production process to identify a sensor observation value that influences the results of quality tests and the like of manufactured goods.
  • a majority of analysis methods, represented by regression analysis, are methods of multidimensionally analyzing observation data on the premise of availability of data that is observed by measurement instruments, such as sensors.
  • PTL 1 describes a method of identifying an influence factor by segmenting data based on nominal scale data when explanatory variables include the nominal scale data and using a multivariate analysis method for each segment.
  • PTL 2 describes a quality variation cause analysis method of a production line, which repeats operation of dividing a plurality of explanatory variables and narrowing down the explanatory variables by performing multiple linear regression analysis for all division groups.
  • NPL 1 describes a method, called L1 regularized logistic regression, which can estimate influence degrees of explanatory variables with high precision when a response variable is a discrete value.
  • NPL 2 describes a random forest classifier that is a classifier implemented using a plurality of decision trees. The techniques described in PTL 1 and 2 and NPL 1 and 2 are also used in factor analysis.
  • Data observed in a production process or the like often includes observation values that vary frequently due to factors such as noise. Thus, data analysis is often difficult when a factor analysis method that uses such as the multivariate analysis described in PTL 1 is used as is for the observation data.
  • preprocessing such as, smoothing data by moving average, is performed on the observation values.
  • the desired is a method with which even those who do not know what is appropriate preprocessing for the observation values can analyze observation values by appropriate preprocessing.
  • the objective of the present invention is to provide a factor analysis device, a factor analysis method, and a factor analysis program that elucidate appropriate preprocessing to be applied to an explanatory time series of an analysis subject and identify the explanatory time series relating to a change in value of a response time series.
  • a factor analysis device includes: a feature extraction unit that extracts feature quantities from an explanatory time series; a feature conversion unit that converts the feature quantities into a feature time series; a feature-time-series influence-degree computation unit that computes, from the feature time series and a response time series, an influence degree of the feature time series on a change in value of the response time series; and an explanatory-time-series influence-degree computation unit that computes, based on the influence degree, an influence degree of the explanatory time series on a change in value of the response time series.
  • a factor analysis method includes: extracting feature quantities from an explanatory time series; converting the feature quantities into a feature time series; computing, from the feature time series and a response time series, an influence degree of the feature time series on a change in value of the response time series; and computing, based on the influence degree, an influence degree of the explanatory time series on a change in value of the response time series.
  • a factor analysis program causes a computer to execute: feature extraction processing of extracting feature quantities from an explanatory time series; feature conversion processing of converting the feature quantities into a feature time series; feature-time-series influence-degree computation processing of computing, from the feature time series and a response time series, an influence degree of the feature time series on a change in value of the response time series; and explanatory-time-series influence-degree computation processing of computing, based on the influence degree, an influence degree of the explanatory time series on a change in value of the response time series.
  • appropriate preprocessing to be applied to an explanatory time series of an analysis subject can be elucidated and an explanatory time series relating to a change in value of a response time series can be identified.
  • FIG. 1 is a block diagram depicting a configuration example of a factor analysis device according to the present invention.
  • FIG. 2 is a flowchart depicting the operation of a factor analysis device 100 .
  • FIG. 3 is an explanatory diagram depicting an example of a method of generating a feature time series from an explanatory time series by a feature-time-series conversion unit 102 .
  • FIG. 4 is a flowchart depicting the operation of the factor analysis device 100 .
  • FIG. 5 is an explanatory diagram depicting an example of explanatory time series and a response time series stored in a time series storage unit 111 .
  • FIG. 6 is an explanatory diagram depicting a generated example of feature time series from each explanatory time series by the feature-time-series conversion unit 102 .
  • FIG. 7 is an explanatory diagram depicting computed examples of influence degrees of feature time series on a response time series by a feature-time-series influence-degree computation unit 1031 using a plurality of multivariate analysis methods.
  • FIG. 8 is an explanatory diagram depicting computed examples of influence degrees of explanatory time series on a response time series by an explanatory-time-series influence-degree computation unit 1032 .
  • FIG. 9 is an explanatory diagram depicting a computed example of influence degrees of an explanatory time series on a response time series by a factor output unit 104 .
  • FIG. 10 is a block diagram depicting main units of a factor analysis device according to the present invention.
  • a factor analysis device is applied to quality management in a manufacturing process.
  • the factor analysis device may be applied to a process other than a manufacturing process or a business other than quality management in a manufacturing process.
  • one kind of response time series of an analysis subject is considered. There may be one or more kinds of response time series of an analysis subject.
  • FIG. 1 is a block diagram depicting a configuration example of a factor analysis device according to the present invention. As depicted in FIG. 1 , the process where the factor analysis device 100 is used in the present exemplary embodiment is linked to a manufacturing process where two or more analysis target devices 200 are used.
  • the analysis target device 200 is a device used in the manufacturing process.
  • the analysis target device 200 measures a plurality of types of measurement values relating to the analysis target device 200 itself at predetermined time intervals, and transmits the measurement values to the factor analysis device 100 .
  • the types of observation values include one or more quality indexes and one or more production conditions of manufactured products.
  • the production conditions include, for example, temperatures, pressures, and gas flow rates.
  • the production conditions are expressed by, for example, numerical values such as integer and decimal.
  • the quality indexes are represented by, for example, numerical values such as integer and decimal.
  • the quality indexes may be represented by a code indicating “abnormal”, “normal” or the like.
  • a “time series” is data where numerical values measured by a sensor are arranged in time order with predetermined time intervals.
  • An “explanatory time series” is a time series that can be obtained by arranging observation values representing production conditions measured by each of the analysis target devices 200 in time order.
  • the explanatory time series widely include production conditions indicating operation conditions of a device, such as adjustment values, temperatures, pressures, gas flow rates, and voltages of the device.
  • response time series is a time series that can be obtained by arranging observation values representing quality indexes measured by each of the analysis target devices 200 in time order. While one kind of response time series of an analysis subject is considered in the present exemplary embodiment, response time series may widely include evaluation indexes of manufactured products or the like, which can be obtained when a device is operated under the production conditions represented by explanatory time series, such as quality and yields.
  • the factor analysis device 100 depicted in FIG. 1 includes an observation data collection unit 101 , a feature-time-series conversion unit 102 , an influence degree computation unit 103 , a factor output unit 104 , a time series storage unit 111 , a feature-time-series storage unit 112 , and an influence degree storage unit 113 .
  • the observation data collection unit 101 has a function of obtaining observation values from the analysis target device 200 .
  • the observation data collection unit 101 stores the obtained observation values in the time series storage unit 111 .
  • the time series storage unit 111 has a function of storing the observation values obtained by the observation data collection unit 101 as time series data.
  • the time series storage unit 111 includes an explanatory-time-series storage unit 1111 and a response-time-series storage unit 1112 .
  • the explanatory-time-series storage unit 1111 stores observation values relating to the production conditions from among the observation values obtained by the observation data collection unit 101 as explanatory time series.
  • the response-time-series storage unit 1112 stores observation values relating to the quality indexes from among the observation values obtained by the observation data collection unit 101 as response time series.
  • the feature-time-series conversion unit 102 has a function of retrieving explanatory time series from the explanatory-time-series storage unit 1111 and converting feature quantities extracted from the explanatory time series into a feature time series.
  • the feature-time-series conversion unit 102 includes a feature extraction unit 1021 and a feature conversion unit 1022 .
  • the feature extraction unit 1021 retrieves a partial time series that is a predetermined time range portion of an explanatory time series from the explanatory-time-series storage unit 1111 , and extracts a feature quantity from the retrieved partial time series.
  • the details of a method of retrieving a partial time series and a method of extracting a feature quantity will be described later.
  • the feature conversion unit 1022 converts the feature quantities extracted by the feature extraction unit 1021 into a feature time series by arranging the feature quantities in time order.
  • the feature conversion unit 1022 stores the generated feature time series in the feature-time-series storage unit 112 .
  • the feature-time-series storage unit 112 has a function of storing the feature time series generated by the feature-time-series conversion unit 102 .
  • the influence degree computation unit 103 has a function of retrieving feature time series from the feature-time-series storage unit 112 and a response time series from the response-time-series storage unit 1112 , and computing an influence degree of the explanatory time series on a change in value of the response time series based on the retrieved data.
  • the influence degree computation unit 103 includes a feature-time-series influence-degree computation unit 1031 and an explanatory-time-series influence-degree computation unit 1032 .
  • the feature-time-series influence-degree computation unit 1031 retrieves feature time series from the feature-time-series storage unit 112 and a response time series from the response-time-series storage unit 1112 .
  • the feature-time-series influence-degree computation unit 1031 computes, for each of the retrieved feature time series by using one or more multivariate analysis methods, influence degrees of the feature time series on the response time series.
  • the number of the computed influence degrees is as much as the number of the used multivariate analysis methods for each of the retrieved feature time series. The details of a method of computing an influence degree of a feature-time-series will be described later.
  • the feature-time-series influence-degree computation unit 1031 stores the computed influence degrees of the feature time series in a feature-time-series influence-degree storage unit 1131 .
  • the explanatory-time-series influence-degree computation unit 1032 retrieves the influence degrees of feature time series each correlated with the one or more multivariate analysis methods from the feature-time-series influence-degree storage unit 1131 .
  • the explanatory-time-series influence-degree computation unit 1032 computes an influence degree of an explanatory time series on the response time series from the retrieved influence degrees of the feature time series based on the information of the explanatory time series as the extraction source of the feature quantities. The details of a method of computing an influence degree of an explanatory time series will be described later.
  • the explanatory-time-series influence-degree computation unit 1032 stores the computed influence degree of the explanatory time series in an explanatory-time-series influence-degree storage unit 1132 .
  • the influence degree storage unit 113 has a function of storing the influence degrees of the feature time series and the influence degrees of the explanatory time series that are computed by the influence degree computation unit 103 .
  • the influence degree storage unit 113 includes the feature-time-series influence-degree storage unit 1131 and the explanatory-time-series influence-degree storage unit 1132 .
  • the feature-time-series influence-degree storage unit 1131 stores the influence degrees of the feature time series computed by the feature-time-series influence-degree computation unit 1031 .
  • the explanatory-time-series influence-degree storage unit 1132 stores the influence degrees of the explanatory time series computed by the explanatory-time-series influence-degree computation unit 1032 .
  • the factor output unit 104 has a function of retrieving the influence degrees of explanatory time series from the explanatory-time-series influence-degree storage unit 1132 in descending order of the influence degrees and outputting explanatory time series corresponding to the retrieved influence degrees of the explanatory time series as factor candidates that influence a change in value of the response time series.
  • the factor output unit 104 has a function of retrieving the influence degrees of feature time series from the feature-time-series influence-degree computation unit 1031 in descending order of the influence degrees and outputting feature quantities corresponding to the retrieved influence degrees of the feature time series as candidates of processing subjects in preprocessing.
  • the factor analysis device 100 in the present exemplary embodiment is implemented, for example, using a Central Processing Unit (CPU) that executes processing in accordance with a program.
  • the factor analysis device 100 may be implemented using a computer that includes a CPU and a recording medium storing a program and operates by control of the CPU in accordance with the program.
  • the observation data collection unit 101 , the feature-time-series conversion unit 102 , the influence degree computation unit 103 , and the factor output unit 104 are implemented, for example, using the CPU that executes processing in accordance with a program control.
  • the time series storage unit 111 , the feature-time-series storage unit 112 , and the influence degree storage unit 113 are implemented, for example, using a Random Access Memory (RAM).
  • the time series storage unit 111 , the feature-time-series storage unit 112 , and the influence degree storage unit 113 may be implemented using one storage medium or a plurality of storage mediums.
  • FIG. 2 is a flowchart depicting the operation of the factor analysis device 100 .
  • the observation data collection unit 101 of the factor analysis device 100 collects sensor observation values from the analysis target devices 200 (step S 101 ).
  • the observation data collection unit 101 determines whether the sensor observation value is an observation value relating to production conditions or an observation value relating to quality indexes (step S 102 ).
  • the observation data collection unit 101 stores the observation value in the explanatory-time-series storage unit 1111 of the time series storage unit 111 (step S 103 ). If the sensor observation value is an observation value relating to quality indexes (NO at step S 102 ), the observation data collection unit 101 stores the observation value in the response-time-series storage unit 1112 (step S 104 ).
  • the observation data collection unit 101 determines whether all sensor observation values is collected from the analysis target devices 200 (step S 105 ). If there is still uncollected sensor observation values (NO at step S 105 ), the observation data collection unit 101 repeats the processing from step S 101 . If all sensor observation values is collected (YES at step S 105 ), the observation data collection unit 101 proceeds the processing to step S 111 .
  • the feature extraction unit 1021 selects one explanatory time series, of which feature quantity is not extracted yet, stored in the explanatory-time-series storage unit 1111 , and retrieves the selected explanatory time series (step S 111 ). Next, the feature extraction unit 1021 arranges the left end of the window, in which a partial time series as a subject of extracting a feature quantity is set, at a time series start time (step S 112 ).
  • the range of time, for which a time series is retrieved is referred to as a “window.”
  • the feature extraction unit 1021 extracts a feature quantity from the partial time series within the range of the window (step S 113 ).
  • the feature extraction unit 1021 determines whether the right end of the window reaches the end time of the explanatory time series (step S 114 ). If not reached (NO at step S 114 ), the feature extraction unit 1021 moves the window by one time point to right, that is, toward the end time (step S 115 ). After moving the window, the feature extraction unit 1021 returns to the processing of step S 113 . The feature extraction unit 1021 repeats the processing of steps S 113 to S 115 until the right end of the window reaches the end time of the explanatory time series.
  • the feature conversion unit 1022 converts feature quantities extracted by the feature extraction unit 1021 into a feature time series by arranging the feature quantities in time order. Then, the feature conversion unit 1022 stores the generated feature time series in the feature-time-series storage unit 112 (step S 116 ).
  • the feature-time-series conversion unit 102 determines whether feature time series is generated from all the explanatory time series stored in the explanatory-time-series storage unit 1111 (step S 117 ). When there is an explanatory time series, of which feature time series is not generated (NO at step S 117 ), the feature-time-series conversion unit 102 repeats the processing of steps S 111 to S 116 . When feature time series is generated from all the explanatory time series (YES at step S 117 ), the feature-time-series conversion unit 102 proceeds to the processing of step S 121 .
  • the feature-time-series influence-degree computation unit 1031 retrieves a response time series from the response-time-series storage unit 1112 and feature time series from the feature-time-series storage unit 112 respectively (step S 121 ).
  • the feature-time-series influence-degree computation unit 1031 computes influence degrees of the feature time series on the response time series using one or more multivariate analysis methods (step S 122 ).
  • the feature-time-series influence-degree computation unit 1031 stores the computed influence degree of the feature time series in the feature-time-series influence-degree storage unit 1131 .
  • the explanatory-time-series influence-degree computation unit 1032 retrieves the influence degrees of feature time series, which are influence degree of the feature time series on the response time series, from the feature-time-series influence-degree storage unit 1131 . Then, the explanatory-time-series influence-degree computation unit 1032 computes an influence degree of the explanatory time series on the response time series based on the information of the explanatory time series as the extraction source of the feature quantities (step S 123 ). The explanatory-time-series influence-degree computation unit 1032 stores the computed influence degree of the explanatory time series in the explanatory-time-series influence-degree storage unit 1132 .
  • the factor output unit 104 integrates the results of the influence degrees of the explanatory time series stored in the explanatory-time-series influence-degree storage unit 1132 .
  • the factor output unit 104 outputs an explanatory time series, of which integrated influence degree is high, as an influence factor (step S 124 ). In this way, the factor analysis device 100 according to the present exemplary embodiment ends the processing.
  • While one response time series is assumed as an analysis subject in the present exemplary embodiment, there may be a plurality of response time series of the analysis subject. If there are a plurality of response time series, the factor analysis device 100 can identify respective explanatory time series that influence the response time series by performing the processing of FIG. 2 for each of the response time series separately.
  • the feature extraction unit 1021 moves a window used for extracting a feature quantity from the explanatory time series to right by one time point
  • the feature extraction unit 1021 may move the window to right by two or more time points at once. If the window is shifted to right by t time points at once, feature quantities at (T ⁇ w)/t points are extracted from one explanatory time series.
  • T is the number of all the time points
  • w is the number of time points for reading
  • t is the number of time points for shifting.
  • the feature extraction unit 1021 may use any kind of feature quantity for feature quantities extracted from the explanatory time series.
  • the feature extraction unit 1021 can use, as a feature quantity to be extracted, basic statistics, such as average and variance, an autoregression coefficient, a frequency distribution, a correlation coefficient with other partial time series, and the like.
  • the basic statistics is computed from a value at each time point in a partial time series.
  • the autoregression coefficient is computed by using an autoregression model for a partial time series and fitting the partial time series by a least squares method or the like.
  • the frequency distribution is computed by performing Fast Fourier Transform (FFT) on a partial time series and extracting frequency components.
  • FFT Fast Fourier Transform
  • the correlation coefficient with other partial time series is computed by calculating a correlation coefficient with a partial time series that is cut out from other explanatory time series in the corresponding window.
  • the feature-time-series influence-degree computation unit 1031 may use any kind of method as a multivariate analysis method, as long as the method computes an influence degree of an explanatory variable on a change in a value of a response variable.
  • the feature-time-series influence-degree computation unit 1031 may convert the code into a correlated numerical value. For example, if a response variable indicates “normal” and “abnormal,” by substituting 1 for “normal” and 0 for “abnormal,” the feature-time-series influence-degree computation unit 1031 can use L1 logistic regression described in NPL 1 or a random forest classifier described in NPL 2 as a multivariate analysis method.
  • the analysis target system may be other system as long as the system can acquire system operation information and performance indexes corresponding to the system operation information.
  • the analysis target system may be an IT system, a plant system, a structure, or transportation equipment.
  • a use rate and a use amount of computer resources such as a CPU use rate, a memory use rate, and a disk access frequency, and a use rate and a use amount of communication network resources are used as operation information.
  • performance indexes a consumption power amount and the number of times of arithmetic operations are used.
  • FIGS. 3 and 5 to 9 are numerical computation results based on actually performed operation.
  • the configuration of the factor analysis device 100 in the present example is the same configuration as depicted in FIG. 1 .
  • the process where the factor analysis device 100 is used in the present example is linked to a manufacturing process where two or more analysis target devices 200 are used.
  • the analysis target device 200 is a device that is used in a manufacturing process.
  • the factor analysis device 100 includes the observation data collection unit 101 , the feature-time-series conversion unit 102 , the influence degree computation unit 103 , the factor output unit 104 , the time series storage unit 111 , the feature-time-series storage unit 112 , and the influence degree storage unit 113 .
  • the influence degree computation unit 103 includes the feature-time-series influence-degree computation unit 1031 and the explanatory-time-series influence-degree computation unit 1032 .
  • the time series storage unit 111 includes the explanatory-time-series storage unit 1111 and the response-time-series storage unit 1112 .
  • the influence degree storage unit 113 includes the feature-time-series influence-degree storage unit 1131 and the explanatory-time-series influence-degree storage unit 1132 .
  • FIG. 3 is an explanatory diagram depicting an example of a method of generating a feature time series from an explanatory time series by a feature-time-series conversion unit 102 .
  • FIG. 3 depicts an explanatory time series as a subject, partial time series as portions of the explanatory time series, and a feature time series that is generated based on feature quantities extracted from the partial time series.
  • the feature extraction unit 1021 retrieves an explanatory time series stored in the explanatory-time-series storage unit 1111 . Next, the feature extraction unit 1021 retrieves data of a w time point portion from the start point from among the retrieved explanatory time series, and obtains a partial time series.
  • the feature extraction unit 1021 extracts one or more kinds of feature quantities from a partial time series corresponding to the retrieved window, and obtains the one or more kinds of feature quantities of real number values.
  • the feature extraction unit 1021 may use a statistic amount, such as average and variance, an autoregression coefficient, a frequency distribution, a correlation coefficient with other explanatory time series, and the like.
  • the feature extraction unit 1021 moves the window by one time point portion ahead (right) and repeats the processing of feature extraction until the right end of the window reaches the end point.
  • n is the number of kinds of feature quantities used.
  • the feature conversion unit 1022 converts the feature quantities into a feature time series by arranging the feature quantities of T-w real number values obtained by the feature extraction unit 1021 in time order. As is obvious from the conversion operation by the feature conversion unit 1022 , in particular, if the feature quantities are converted into an average feature time series, the feature time series coincides with a time series obtained by moving average of the explanatory time series with width w.
  • the feature-time-series conversion unit 102 performs the above operation of the feature extraction unit 1021 and the operation of the feature conversion unit 1022 on all explanatory time series stored in the explanatory-time-series storage unit 1111 , and obtains m ⁇ n feature time series.
  • m is the number of explanatory time series that are generation sources of the feature time series.
  • All the generated feature time series are labelled in a manner that enables recognition of explanatory time series as generation sources and the kinds of extracted features.
  • a feature time series which is obtained by extracting feature quantities labelled as “a” from an explanatory time series labelled as “1,” is labelled as “a::1” or the like.
  • an influence degree of the feature time series assumed to be input on the response time series assumed to be output can be computed based on the input and output relationship.
  • the feature-time-series influence-degree computation unit 1031 computes a plurality of influence degrees of one feature time series using a plurality of multivariate analysis methods.
  • the feature quantity is labelled in a form of, for example, “(name of feature quantity)::(name of explanatory time series).”
  • an influence degree is normalized so that the greatest value becomes 1 and the least value becomes 0.
  • the explanatory-time-series influence-degree computation unit 1032 computes an influence degree of an explanatory time series on a response time series based on the influence degree of a feature time series computed by the feature-time-series influence-degree computation unit 1031 .
  • the explanatory-time-series influence-degree computation unit 1032 sums the influence degrees of the feature time series for the multivariate analysis methods used and the explanatory time series that are sources of extraction of the feature quantities separately.
  • the subjects to be summed may be all feature quantities, and may be only some of the feature quantities with higher influence degrees.
  • the factor output unit 104 integrates the influence degrees of the explanatory time series on the response time series, which are computed by a plurality of multivariate analysis methods.
  • the factor output unit 104 sums the influence degrees computed by a plurality of multivariate analysis methods for the explanatory time series separately.
  • the method of summing may be simple summing or summing by weighting methods independently.
  • FIG. 4 is a flowchart depicting the operation of the factor analysis device 100 .
  • the observation data collection unit 101 of the factor analysis device 100 collects 51 kinds of sensor observation values including sensor observation values that represent 50 kinds of production conditions and a sensor observation value that represents one kind of quality index from 51 analysis target devices 200 (step S 51 ).
  • the observation data collection unit 101 generates 50 explanatory time series by separately arranging sensor observation values that represent 50 kinds of production conditions in time order. After generating the 50 explanatory time series, the observation data collection unit 101 stores the 50 explanatory time series in the explanatory-time-series storage unit 1111 .
  • the observation data collection unit 101 generates one response time series by arranging the sensor observation values that represent the quality index in time order. After generating the one response time series, the observation data collection unit 101 stores the one response time series in the response-time-series storage unit 1112 (step S 52 ).
  • FIG. 5 is an explanatory diagram depicting an example of the explanatory time series and the response time series stored in the time series storage unit 111 .
  • FIG. 5 depicts four explanatory time series which are labelled as “1,” “13,” “37” or “50” from among the 50 explanatory time series, and one response time series in the present example.
  • explanatory time series labelled as “13” and the explanatory time series labelled as “37” in FIG. 5 are explanatory time series that are influence factors on the response time series among the 50 explanatory time series. The fact that these two explanatory time series are the influence factors is unknown for a user at the time when the processing of step S 52 is performed.
  • the feature-time-series conversion unit 102 generates feature time series from all the explanatory time series (step S 53 ).
  • the feature-time-series conversion unit 102 generates a plurality of feature time series from each of the explanatory time series.
  • the feature-time-series conversion unit 102 stores the generated feature time series in the feature-time-series storage unit 112 .
  • FIG. 6 is an explanatory diagram depicting a generated example of the feature time series from each of the explanatory time series by the feature-time-series conversion unit 102 .
  • FIG. 6 depicts feature time series relating to a feature quantity a and a feature quantity b which are extracted from each of the four explanatory time series that are labelled as “1,” “13,” “37,” and “50,” respectively, and the response time series in the present example.
  • the feature time series depicted in FIG. 6 are labelled in a manner that enables recognition of the explanatory time series that are sources and the kinds of extracted feature quantities.
  • the feature time series that is obtained by extracting the feature quantity labelled as “a” from the explanatory time series labelled as “1” is labelled as “a::1.”
  • the feature-time-series influence-degree computation unit 1031 retrieves the response time series from the response-time-series storage unit 1112 and the feature time series from the feature-time-series storage unit 112 (step S 54 ). Next, the feature-time-series influence-degree computation unit 1031 computes influence degrees of the feature time series on the response time series by using one or more multivariate analysis methods (step S 55 ). The feature-time-series influence-degree computation unit 1031 stores the computed influence degrees of the feature time series in the feature-time-series influence-degree storage unit 1131 .
  • FIG. 7 is an explanatory diagram depicting computed examples of the influence degrees of the feature time series on the response time series by the feature-time-series influence-degree computation unit 1031 by using a plurality of multivariate analysis methods.
  • FIG. 7 depicts influence degrees in the present example, which are computed separately by multivariate analysis methods of method I, method II, and method III.
  • FIG. 7 depicts the feature time series in descending order of the computed influence degrees.
  • the feature quantity a is determined as an important feature quantity since it is understood that the influence degree on a change of the value of the response time series is large when a is extracted as a feature quantity.
  • the explanatory-time-series influence-degree computation unit 1032 retrieves the influence degrees of the feature time series on the response time series from the feature-time-series influence-degree storage unit 1131 . Then, the explanatory-time-series influence-degree computation unit 1032 computes the influence degrees of the explanatory time series based on the information of the explanatory time series that are the source of extracting the feature quantities (step S 56 ). The explanatory-time-series influence-degree computation unit 1032 stores the computed influence degrees of the explanatory-time-series in the explanatory-time-series influence-degree storage unit 1132 .
  • FIG. 8 is an explanatory diagram depicting computed examples of the influence degrees of the explanatory time series on the response time series by the explanatory-time-series influence-degree computation unit 1032 .
  • FIG. 8 depicts the influence degrees computed by each of the multivariate analysis methods of method I, method II, and method III in the present example.
  • FIG. 8 depicts the explanatory time series in descending order of the computed influence degrees.
  • the factor output unit 104 integrates the computed influence degrees of the explanatory time series, and outputs the explanatory time series with the high integrated influence degree as an influence factor (step S 57 ).
  • the factor output unit 104 outputs an important feature quantity from the computation results of the influence degrees of the feature time series. In this way, the factor analysis device 100 according to the present example ends the processing.
  • FIG. 9 is an explanatory diagram depicting a computed example of the influence degrees of the explanatory time series on a response time series by the factor output unit 104 .
  • the computation results depicted in FIG. 9 are the results that are finally obtained by computation of the factor analysis device 100 in the present example.
  • FIG. 9 depicts influence degrees obtained by integrating the influence degrees computed by each of the methods of method I, method II, and method III.
  • FIG. 9 depicts the explanatory time series in descending order of the integrated influence degrees.
  • the explanatory time series with the largest influence degree is the explanatory time series labelled as “13” and the explanatory time series with the second largest influence degree is the explanatory time series labelled as “37.”
  • the explanatory time series labelled as “13” and explanatory time series labelled as “37” are influence factors, it is understood that the explanatory time series that are strongly relating to a response time series are correctly output in the computation results of the influence degrees as depicted in FIG. 9 .
  • FIG. 7 depicts the feature quantity a as an important feature quantity.
  • the feature quantity a largely changes at the point where the response time series changes.
  • the values largely change in conjunction with a change in the response time series.
  • the feature quantity a is determined as an important feature quantity, from which a change in value of the response time series can be easily detected, thus, it is understood that the feature quantity a should be extracted from the explanatory time series in preprocessing.
  • the factor analysis device 100 in the present example can identify appropriate preprocessing and an explanatory time series relating to a change in value of a response time series on the basis of an explanatory time series data set, which is hard to analyze without preprocessing, and a response time series relating to the explanatory time series.
  • the factor analysis device is a factor analysis device that identifies an explanatory time series relating to a change factor of a response time series of a system based on one or more explanatory time series of the system, which are obtained by observing the state of the system that is a subject by one or more sensors, and the response time series which is explained by the explanatory time series.
  • the factor analysis device can not only identify an explanatory time series that is a factor strongly relating to a change in value of the response time series but also provide information of preprocessing that is appropriate for analysis in identifying the explanatory time series, based on the explanatory time series data that needs appropriate preprocessing and the explanatory time series that are explained by the explanatory time series data.
  • appropriate preprocessing in addition to an explanatory time series relating to a change in value of a response time series, can be identified from an explanatory time series data set that is hard to be analyzed without preprocessing and the response time series relating to the explanatory time series.
  • the reason is that, because the feature quantity for which a large influence degree is computed in an analysis process is an important feature quantity that should be extracted from the explanatory time series in the factor analysis, extracting the feature quantity from the explanatory time series is preprocessing that should be applied to the explanatory time series.
  • FIG. 10 is a block diagram depicting the main units of the factor analysis device according to the present invention.
  • the factor analysis device 100 includes, as main components: the feature extraction unit 1021 that extracts feature quantities from an explanatory time series; the feature conversion unit 1022 that converts the feature quantities into a feature time series; the feature-time-series influence-degree computation unit 1031 that computes influence degrees of the feature time series on a change in value of a response time series from the feature time series and the response time series; and the explanatory-time-series influence-degree computation unit 1032 that computes influence degrees of the explanatory time series on a change in value of the response time series based on the influence degrees of the feature time series.
  • the factor analysis device can elucidate appropriate preprocessing to be applied to explanatory time series of an analysis subject and identify an explanatory time series relating to a change in value of a response time series.
  • the feature extraction unit 1021 may extract a feature quantity from a partial time series, which is a portion of an explanatory time series, within a range of a window with a predetermined time range, and the feature conversion unit 1022 may convert the obtained feature quantities into a feature time series when the feature extraction unit 1021 extracts a feature quantities at positions by shifting the window by a predetermined number of time points from start time to end time of the explanatory time series and the window reaches the end time.
  • the factor analysis device can extract feature quantities from partial time series that are cut out by a window and convert the extracted feature quantities into a time series.
  • the factor analysis device 100 may include a factor output unit (for example, factor output unit 104 ) that outputs feature quantities relating to a feature time series with a large influence degree on a change in value of a response time series and an explanatory time series with a large influence degree on a change in value of the response time series.
  • a factor output unit for example, factor output unit 104
  • the factor analysis device can provide feature quantities that should be extracted from explanatory time series of an analysis subject and information of an explanatory time series relating to a change in value of a response time series.
  • the feature extraction unit 1021 may extract one or more kinds of feature quantities from one or more explanatory time series, and the feature conversion unit 1022 may convert the feature quantities into a plurality of feature time series associated with the kinds of feature quantities.
  • the factor analysis device can prepare many candidates of preprocessing by extracting as many kinds of feature quantities as possible from explanatory time series.
  • the feature-time-series influence-degree computation unit 1031 may compute influence degrees of feature time series on a change in value of a response time series by using one or more multivariate analysis methods.
  • the factor analysis device can obtain a feature quantity related with an explanatory time series that is a factor of a change in value of a response variable and identify an explanatory time series that is a factor from a plurality of viewpoints.
  • the feature-time-series influence-degree computation unit 1031 may use L1 regularized logistic regression as one of the multivariate analysis methods.
  • the feature-time-series influence-degree computation unit 1031 may use a random forest classifier as one of the multivariate analysis methods.
  • the factor analysis device 100 may use any of average, standard deviation, skewness, kurtosis, and p-quartile for a feature quantity.
  • the factor analysis device 100 may use an autoregression model coefficient for a feature quantity.
  • the factor analysis device 100 may use a correlation coefficient with an explanatory time series for a feature quantity.
  • the factor analysis device 100 may use a frequency distribution of an explanatory time series for a feature quantity.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Data Mining & Analysis (AREA)
  • Theoretical Computer Science (AREA)
  • Mathematical Physics (AREA)
  • Pure & Applied Mathematics (AREA)
  • Computational Mathematics (AREA)
  • Mathematical Analysis (AREA)
  • Mathematical Optimization (AREA)
  • General Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Algebra (AREA)
  • Probability & Statistics with Applications (AREA)
  • Operations Research (AREA)
  • Evolutionary Biology (AREA)
  • Databases & Information Systems (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Quality & Reliability (AREA)
  • Automation & Control Theory (AREA)
  • Medical Informatics (AREA)
  • Artificial Intelligence (AREA)
  • Computational Linguistics (AREA)
  • Evolutionary Computation (AREA)
  • Computing Systems (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
  • Testing And Monitoring For Control Systems (AREA)
  • Complex Calculations (AREA)
US15/125,302 2014-03-14 2014-12-15 Factor analysis device, factor analysis method, and factor analysis program Active 2035-11-18 US10496730B2 (en)

Applications Claiming Priority (3)

Application Number Priority Date Filing Date Title
JP2014051096 2014-03-14
JP2014-051096 2014-03-14
PCT/JP2014/006232 WO2015136586A1 (fr) 2014-03-14 2014-12-15 Dispositif d'analyse factorielle, procédé d'analyse factorielle et programme d'analyse factorielle

Publications (2)

Publication Number Publication Date
US20170109324A1 US20170109324A1 (en) 2017-04-20
US10496730B2 true US10496730B2 (en) 2019-12-03

Family

ID=54071067

Family Applications (1)

Application Number Title Priority Date Filing Date
US15/125,302 Active 2035-11-18 US10496730B2 (en) 2014-03-14 2014-12-15 Factor analysis device, factor analysis method, and factor analysis program

Country Status (4)

Country Link
US (1) US10496730B2 (fr)
EP (1) EP3118750A4 (fr)
JP (1) JPWO2015136586A1 (fr)
WO (1) WO2015136586A1 (fr)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190086522A1 (en) * 2017-09-15 2019-03-21 Kabushiki Kaisha Toshiba Distance measuring device

Families Citing this family (13)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2017117089A (ja) * 2015-12-22 2017-06-29 ローム株式会社 センサノード、センサネットワークシステム、および監視方法
US11580414B2 (en) 2016-03-23 2023-02-14 Nec Corporation Factor analysis device, factor analysis method, and storage medium on which program is stored
WO2017168524A1 (fr) * 2016-03-28 2017-10-05 株式会社日立製作所 Dispositif de serveur d'analyse, système d'analyse de données et procédé d'analyse de données
JP6702081B2 (ja) * 2016-08-15 2020-05-27 富士通株式会社 判定装置、判定方法、および判定プログラム
JP6835098B2 (ja) * 2016-11-28 2021-02-24 日本電気株式会社 要因分析方法、要因分析装置および要因分析プログラム
WO2019013196A1 (fr) * 2017-07-14 2019-01-17 パナソニックＩｐマネジメント株式会社 Dispositif de gestion de fabrication, système de fabrication, et procédé de gestion de fabrication
JP7139625B2 (ja) * 2017-08-04 2022-09-21 富士電機株式会社 要因分析システム、要因分析方法およびプログラム
WO2019142346A1 (fr) * 2018-01-22 2019-07-25 日本電気株式会社 Système d'analyse, procédé d'analyse et support d'enregistrement
US20200342048A1 (en) * 2018-01-22 2020-10-29 Nec Corporation Analysis device, analysis method, and recording medium
KR102109369B1 (ko) * 2018-03-16 2020-05-28 울산과학기술원 시계열 데이터의 변화를 예측하고 그 이유를 설명하는 인공지능 시스템
DE102018109816B3 (de) * 2018-04-24 2019-10-24 Yxlon International Gmbh Verfahren zur Gewinnung mindestens eines signifikanten Merkmals in einer Serie von Bauteilen gleichen Typs und Verfahren zur Klassifikation eines Bauteils eienr solchen Serie
JP7056497B2 (ja) * 2018-10-03 2022-04-19 トヨタ自動車株式会社 重回帰分析装置及び重回帰分析方法
JP7368189B2 (ja) * 2019-11-07 2023-10-24 ファナック株式会社 分析装置

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2002110493A (ja) 2000-10-04 2002-04-12 Matsushita Electric Ind Co Ltd 加工プロセス工程の異常抽出方法及び装置
US20050209820A1 (en) * 2004-03-10 2005-09-22 International Business Machines Corporation Diagnostic data detection and control
US20060015263A1 (en) * 2004-07-10 2006-01-19 Stupp Steven E Apparatus for determining association variables
JP2006318263A (ja) 2005-05-13 2006-11-24 Sharp Corp 情報分析システム、情報分析方法及びプログラム
JP2009258890A (ja) 2008-04-15 2009-11-05 Sharp Corp 影響要因特定装置
JP2009301341A (ja) 2008-06-13 2009-12-24 Nec Corp サービシステム、サービスシステム管理方法、及びプログラム
US20110112998A1 (en) 2009-11-11 2011-05-12 International Business Machines Corporation Methods and systems for variable group selection and temporal causal modeling
US20120083917A1 (en) * 2009-06-10 2012-04-05 Fisher-Rosemount Systems, Inc. Predicted fault analysis
US20130310660A1 (en) * 2007-11-14 2013-11-21 Medasense Biometrics Ltd. System and method for pain monitoring using a multidimensional analysis of physiological signals

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP2002110493A (ja) 2000-10-04 2002-04-12 Matsushita Electric Ind Co Ltd 加工プロセス工程の異常抽出方法及び装置
US20050209820A1 (en) * 2004-03-10 2005-09-22 International Business Machines Corporation Diagnostic data detection and control
US20060015263A1 (en) * 2004-07-10 2006-01-19 Stupp Steven E Apparatus for determining association variables
JP2006318263A (ja) 2005-05-13 2006-11-24 Sharp Corp 情報分析システム、情報分析方法及びプログラム
US20130310660A1 (en) * 2007-11-14 2013-11-21 Medasense Biometrics Ltd. System and method for pain monitoring using a multidimensional analysis of physiological signals
JP2009258890A (ja) 2008-04-15 2009-11-05 Sharp Corp 影響要因特定装置
JP2009301341A (ja) 2008-06-13 2009-12-24 Nec Corp サービシステム、サービスシステム管理方法、及びプログラム
US20120083917A1 (en) * 2009-06-10 2012-04-05 Fisher-Rosemount Systems, Inc. Predicted fault analysis
US20110112998A1 (en) 2009-11-11 2011-05-12 International Business Machines Corporation Methods and systems for variable group selection and temporal causal modeling

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
Eiji Motohashi, et al. An Analysis of Purchase History Data by a Brand Choice Model Considering the Change of Market Structure, Journal of Marketing Science, Mar. 28, 2013, vol. 21 No. 1, pp. 37-59.
International Search Report corresponding to PCT/JP2014/006232 dated Feb. 10, 2015 (one page).
ISBN: 1-58113-838-5, Andrew Y. Ng, "Feature selection, L1 vs. L2 regularization, and rotational invariance," in Proceedings of the 21st International Conference of Machine Learning, pp. 78-85, 2004 (8 pages).
ISSN: 0885-6125, Breiman. L, "Random Forests," Machine Learning, vol. 45, No. 1, pp. 5-32, 2001.

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20190086522A1 (en) * 2017-09-15 2019-03-21 Kabushiki Kaisha Toshiba Distance measuring device
US10866311B2 (en) * 2017-09-15 2020-12-15 Kabushiki Kaisha Toshiba Distance measuring device

Also Published As

Publication number Publication date
US20170109324A1 (en) 2017-04-20
WO2015136586A1 (fr) 2015-09-17
EP3118750A1 (fr) 2017-01-18
JPWO2015136586A1 (ja) 2017-04-06
EP3118750A4 (fr) 2017-11-08

Similar Documents

Publication Publication Date Title
US10496730B2 (en) Factor analysis device, factor analysis method, and factor analysis program
US11216741B2 (en) Analysis apparatus, analysis method, and non-transitory computer readable medium
JP6243080B1 (ja) プリプロセッサおよび異常予兆診断システム
US9658916B2 (en) System analysis device, system analysis method and system analysis program
JP6594044B2 (ja) 実時系列内の異常を検出するための方法
Sarda-Espinosa et al. Conditional inference trees for knowledge extraction from motor health condition data
EP2963553B1 (fr) Dispositif et procédé d'analyse de systèmes
US20200341454A1 (en) Factor analysis method, factor analysis device, and factor analysis program
WO2016079972A1 (fr) Appareil d'analyse de facteur, procédé d'analyse de facteur et support d'enregistrement et système d'analyse de facteur
Dobos et al. On-line detection of homogeneous operation ranges by dynamic principal component analysis based time-series segmentation
US11216534B2 (en) Apparatus, system, and method of covariance estimation based on data missing rate for information processing
JP6631540B2 (ja) 情報処理システム、変化点検出方法、およびプログラム
Cholewa et al. Estimation of the number of states for gesture recognition with Hidden Markov Models based on the number of critical points in time sequence
CN112966222A (zh) 一种时间序列异常数据检测方法及相关设备
JPWO2014132611A1 (ja) システム分析装置、及び、システム分析方法
US11989013B2 (en) Abnormality detection apparatus, abnormality detection system, and learning apparatus, and methods for the same and non-temporary computer-readable medium storing the same
EP4160341A1 (fr) Dispositif, procédé et programme d'identification de cause de modulation anormale
Ye et al. Scaling up discrete distribution clustering using ADMM
US20190180180A1 (en) Information processing system, information processing method, and recording medium
CN116482597A (zh) 电能表运行数据变量筛选方法、电子设备及存储介质
CN106909497B (zh) 测试方法和装置
US20200342048A1 (en) Analysis device, analysis method, and recording medium
US20180046950A1 (en) Time-series-data processing device
Cupek et al. Improving KPI based performance analysis in discrete, multi-variant production
US11580197B2 (en) Analysis system, analysis method, and recording medium

Legal Events

Date Code Title Description
AS Assignment

Owner name: NEC CORPORATION, JAPAN

Free format text: ASSIGNMENT OF ASSIGNORS INTEREST;ASSIGNOR:MIZOGUCHI, TAKEHIKO;REEL/FRAME:039701/0178

Effective date: 20160802

STPP Information on status: patent application and granting procedure in general

Free format text: FINAL REJECTION MAILED

STPP Information on status: patent application and granting procedure in general

Free format text: RESPONSE AFTER FINAL ACTION FORWARDED TO EXAMINER

STPP Information on status: patent application and granting procedure in general

Free format text: NOTICE OF ALLOWANCE MAILED -- APPLICATION RECEIVED IN OFFICE OF PUBLICATIONS

STPP Information on status: patent application and granting procedure in general

Free format text: PUBLICATIONS -- ISSUE FEE PAYMENT VERIFIED

STCF Information on status: patent grant

Free format text: PATENTED CASE

MAFP Maintenance fee payment

Free format text: PAYMENT OF MAINTENANCE FEE, 4TH YEAR, LARGE ENTITY (ORIGINAL EVENT CODE: M1551); ENTITY STATUS OF PATENT OWNER: LARGE ENTITY

Year of fee payment: 4